The Product Desk

What Changed This Week

An engineer opened Cursor on Tuesday and ran the same Claude-backed agent loop she has run every day for six months. Starting June 15, that loop costs roughly ten times what it cost her on Monday. Anthropic announced that every Claude subscription now includes API credits equal to the plan's dollar amount ($200 plan = $200 API credits). Pitched as generous. For the cohort using Claude through third-party harnesses — Cursor, Cline, OpenCode, T3 Code, Zed — at effective 70-90% discounts to API pricing, it is a price increase of roughly an order of magnitude. Third-party tool usage gets a separate credit pool, and once those credits burn, the meter runs at full API rates.

OpenAI responded within hours. Sam Altman offered 2 months of free Codex to enterprise customers who switch within 30 days. That is displacement pricing timed to a moment of developer frustration. Criticism came immediately from Theo, Jeremy Howard, Matt Pocock, and Omar Sanseviero.

Why This Is Happening Now

Anthropic hired a CFO and is likely targeting an October 2026 IPO. The previous model — power users getting enormous implicit subsidies through third-party tools — does not produce the revenue-per-user metrics public market investors read on the first page of an S-1. Ramp data confirms Anthropic now leads business adoption at 34.4% vs OpenAI's 32.3%, which reduces the incentive to subsidize. Expect at least one more pricing adjustment before October.

The era of subsidized AI inference through integrations is ending. Model the cost impact now — not after the credit pool runs dry.

The Decision Framework

Draw a 2x2 before next sprint planning. Axis one: is Claude usage load-bearing for a production workflow, or exploratory? Axis two: is the harness replaceable with Anthropic-native tooling at similar quality, or not?

• Load-bearing + replaceable: renegotiate with Anthropic inside the 30-day window while leverage is real.
• Load-bearing + not replaceable: pilot Codex on the 2-months-free offer this week.
• Exploratory in either cell: stop paying metered rates for exploration. Move to whichever vendor is currently subsidizing it.

Contradiction Worth Surfacing

Sources disagree on whether this benefits OpenAI. Vercel's production data shows 61% of AI spend goes to Anthropic (Opus for reasoning) while Google captures 38% of token volume (Flash for cheap tasks). Teams are already routing across models in production. The pricing shock probably accelerates model abstraction rather than handing share to OpenAI. The 50% rate limit increase Anthropic is offering for two months is a grace period, not a concession.

The Pattern Nobody Is Tracking

Kellie Romack, ServiceNow's CDIO, opened her Anthropic billing dashboard and watched the full-year Anthropic budget get consumed before mid-2026. She cannot name which users drove the spend, or which workloads. Anthropic does not ship the telemetry that would answer those questions. PagerDuty and National Life Group describe the same hole. Nimesh Mehta at National Life calls Anthropic 'great for consumer usage but not great for companies.'

This is not a ServiceNow problem. It is what happens when AI features ship without per-user, per-feature inference cost visibility. A handful of power users consume most of the budget. Finance discovers it at the quarterly review, after the money is gone.

AI costs are structurally unpredictable. The model providers have not built the instrumentation customers need to govern them.

Why Costs Won't Fall This Year

Three data points that invalidate the 'inference costs decline quarterly' assumption:

1. Nebius reports 4+ customers competing for every GPU brought online. Demand is ahead of supply for the foreseeable horizon.
2. OpenAI committed $20B to Cerebras in a multi-year capacity lockup. Not a temporary shortage.
3. Anthropic is raising at $900B and customers are absorbing price increases without churning. Pricing power is consolidating.

The compute crunch is not resolving. Microsoft is spending $100B+ on OpenAI infrastructure. Tencent plans to 'significantly' increase H2 2026 spending. When the largest buyers are capacity-constrained, the smaller buyers are not getting a better quote next quarter.

The Tokenmaxxing Problem

A second failure mode shows up next to budget overruns: 'tokenmaxxing'. Employees use AI tools heavily with no productivity correlation. Amazon mandated 80%+ AI tool usage and staff are gaming token leaderboards. What teams report (tokens consumed, sessions per week) shows adoption. What they do not report (outcomes per token, task completion rate) shows whether adoption produced value. Adoption is what gets pitched. Value is what gets done.

What ServiceNow Built to Fix It

ServiceNow built AI Control Tower internally and staffed a dedicated person to watch Anthropic consumption. They now sell the tool to their own customers. The diagnostic for product teams shipping AI features is simple. On one axis: can a customer's finance team see per-user spend by Friday, or does the customer have to build that themselves. On the other: do reports tie tokens to outcomes, or only to usage. AI cost governance is moving from nice-to-have to procurement blocker. The teams that ship the instrumentation in the next two quarters get the renewal conversation. The teams that ship more features get the audit.

SAP, ServiceNow, and Notion All Shipped Agent-Callable Surfaces This Week

SAP shipped a Knowledge Graph for agent context alongside a €100M partner fund for its Autonomous Enterprise initiative. ServiceNow launched Action Fabric, decoupling workflow logic from UI and exposing it through MCP servers for third-party agent execution. Notion shipped an External Agents API that lets Claude, Codex, Cursor, and Devin operate directly inside Notion. Three of the largest enterprise vendors landing in the same place in the same week is a commitment to letting someone else's agent drive the product on behalf of the buyer.

The buyer behavior is already showing up in demos. A Fortune 500 procurement manager is asking vendors: 'Can our agents call this directly, or do my people have to click through your UI?' Two vendors couldn't answer; the one that could moved to the next round.

The enterprise buying committee has moved from 'show me the dashboard' to 'can our agents orchestrate your workflows.' Products without agent-callable APIs face quiet removal from shortlists within 2-3 quarters.

What 'Agent-Callable' Actually Means

This is not 'build an agent.' It is narrower: can the product's core workflows be invoked by an agent that is not yours, without a human clicking through the UI, by Q4? The build is smaller than most decks suggest: a week of scoping, 2-4 weeks of build for an MCP server against an existing API. The decision that takes longer is whether the UI should be restructured around the assumption that an agent is the primary first-touch user for a growing share of sessions.

The Security Implication Nobody Is Pricing In

Legacy bot detection has an 81% AI agent bypass rate. A product that relies on CAPTCHA, rate limiting, or behavioral analysis for access control should be treated as compromised. Agent-callable APIs need their own authentication model: scoped credentials, audit trails, and spend caps. The OpenClaw incident, where agents deleted all user emails, is what happens when agent access ships without principle-of-least-privilege design.

The Forcing Function

Pull the last 20 support tickets and feature requests from top-decile accounts. Count how many assume a human in the seat versus an agent or integration doing the work. If the ratio has moved even 10 points toward agents in the last two quarters, the headless layer is a retention bet tied to the next renewal cycle, not a platform bet tied to the next board meeting.

The Lemkin Case Study Changes the Math

Jason Lemkin cut his Salesforce seat count from 10+ humans to 2 humans plus 1 API seat. His bill went from $12K to $22K — up 83%. The vendor kept the revenue. The vendor also lost nine daily active users. Those two facts live in the same account, and they describe where every B2B SaaS product's economics are heading.

a16z's May 2026 GTM thesis calls this the shift from systems of record to systems of intelligence. CRM usage has actually risen since AI tool adoption — but it's being consumed as infrastructure at the API layer, not as the decision surface. The rep opens the CRM to log. She opens a different tool to decide who to call. The intelligence layer that reads notes, weighs pipeline, and recommends actions is increasingly not the CRM vendor's product.

Switching costs are migrating from 'our data lives in Salesforce' to 'our workflows, reasoning, and institutional context live in our AI layer' — creating a new moat category entirely.

What This Means for Pricing Strategy

If pricing is still per-seat, you are optimizing for a shrinking metric. The Lemkin account shows the pattern: seats compress, API/workflow volume expands, total spend grows. Three implications for your pricing page:

• Prototype consumption/outcome-based pricing tier alongside seats — test willingness to pay for agent API access and per-action pricing
• Measure whether customers expand seats when agents deploy — if they do not, the 'agents multiply seat value' narrative collapses
• Define what 'institutional context' your product accumulates — playbooks, decision history, rules the agent executes against. That context is the new switching cost.

The TAM Reframe

Software historically captured only 5-10% of GTM spend — the other 90-95% was payroll. AI agents don't replace payroll. They make each payroll dollar more productive. The addressable market is not the ~$150B of existing GTM software. It is a slice of the payroll line that software could never reach before. Teams positioning against Salesforce's $140B are anchored to the wrong number.

The Window Is Narrowing

Salesforce and HubSpot are rapidly building API-first AI offerings. Startups deploying agentic workflows faster than incumbents have a window — but it's closing. The forcing function: pick one workflow with structured inputs and measurable outputs, and go deep enough that institutional context accumulates before the platform play catches up. Depth in one workflow beats breadth across many.